The launch combines a unified, action-oriented world model with Ambient Capture Engine 2.0, three commercial solutions and a new industry benchmark, connecting high-density data with real-world robotic operations.

SHANGHAI, CHINA - Media OutReach Newswire - 19 July 2026 - ACE ROBOTICS today unveiled Kairos 3.1, its latest action-oriented world model, alongside Ambient Capture Engine 2.0 and three commercial solutions spanning instant retail, hospitality and outdoor service scenarios. The announcements were made at a forum ACE ROBOTICS organized during the 2026 World Artificial Intelligence Conference, convened around the theme of advancing physical AI from "understanding" to "execution." The event brought together economists, technology executives and embodied AI researchers to examine how general-purpose world models can move from laboratory research to industrial deployment. The forum also marked the launch of PHYSICAL IQ, a unified benchmark for embodied physical intelligence. It was jointly initiated by the Shanghai Artificial Intelligence Association, the Shenzhen Loop Area Institute and the East China branch of the China Academy of Information and Communications Technology, with participation from more than 20 universities and industry partners. Thomas J. Sargent, Nobel laureate in Economic Sciences, discussed the limitations of current intelligent systems in rare and previously unseen situations. A world model built for the physical world "In the digital world, a model error may result in a flawed image or paragraph. In the physical world, an incorrect action can have real consequences," said Wang Xiaogang, Chairman of ACE ROBOTICS. "Kairos 3.1 is built around a first-principles approach to embodied world models, helping robots act more reliably in complex and uncertain environments, and accelerating the arrival of physical AI's Kairos moment." Kairos 3.1 is designed as a natively unified model that integrates generative, physical and cognitive intelligence within a single architecture, rather than combining separately developed capabilities. Built on a hybrid Transformer architecture with a shared mixed-attention mechanism, it brings visual observations, language instructions, force and tactile signals, and policy trajectories into a unified latent space. This supports an "understand, reason, execute and reflect" loop. The model can break down long-horizon tasks, simulate physical cause and effect across multiple scenarios, rank candidate actions, execute the selected strategy and evaluate the outcome for further adjustment. At the core of its spatial understanding is ACE-BRAIN-0.5. ACE ROBOTICS said it has achieved state-of-the-art results across 12 public evaluations covering spatial understanding, navigation, manipulation and task-progress assessment, among publicly reported models as of July 2026. In a household laundry scenario, for example, a robot can identify spatial relationships between objects, divide a task into more than a dozen steps and verify each stage in real time. If a failure occurs, it can identify the affected step and restart from that point rather than repeating the entire task. For physical generation and reasoning, Kairos-HomeWorld supports whole-home scene generation and object interaction. It is built on 300,000 residential floor plans, 5,000 simulated home environments and 8,700 3D assets covering six categories of physical properties, designed to reflect common residential layouts in China. The model can simulate multiple action trajectories in parallel and rank them based on predicted success and execution cost before sending instructions to a physical robot. In internal testing, the Kairos 3.1 8B model achieved an inference latency of 125 milliseconds on the NVIDIA Jetson Thor platform at BF16 precision. Its in-house KairosRT computing engine supports real-time, on-device inference. Kairos 3.1 also incorporates self-reflective iteration. When an action fails, the robot can evaluate the result and adjust its strategy. In one test, for example, it changed from a three-finger to a four-finger grasp after an unsuccessful attempt and subsequently completed the task. From high-density data to continuous learning ACE ROBOTICS also introduced what it calls the "Information-Density Law" for embodied models: the value of data is determined not only by its volume, but by whether it contains information capable of changing the outcome of an agent's actions. ACE ROBOTICS classifies embodied data across five information-density levels, from L1 to L5. L1 and L2 data support foundational pre-training; L3 and L4 data remain focused on known tasks; and at L5, data incorporates three-dimensional force and tactile signals, failure-recovery trajectories and variables from open environments. The company says these higher-density forms of interaction data support stronger generalization,...